Comment by aenis

Their pace makes me wonder how effectively do frontier labs protect their weights against motivated, capable adversaries. By now the value of something like Mythos is well in double digit billions, and more so in strategic advantage.

Considering the hosting happens across many providers, and many geographies, as does training, I wonder how airtight the CC tech really is to prevent, I don't know, key extraction from the secure enclave of the GPUs. This requires physical access and cutting edge techniques, but this is also the absolute cutting edge of secrets-worth-stealing.

I'd not be surprised if the distillation attacks via API were a smokescreen to theft of actual weights. Long shot, but given the motivations, and the general weird state of US AI labs. I'd not surprise me.